Don't reap idle Claude sessions that still run background subagents #1443

In the Solenta repo (~/code/coder), electron/runner.js scheduleClaudeIdleReap(threadId) retires idle kept-alive Claude CLI sessions in two ways:

• after CLAUDE_IDLE_REAP_MS (30 min);
• immediately for the oldest idle sessions once more than CLAUDE_IDLE_MAX (3) are idle.

Neither path checks whether the session still has background subagents running: thread.subagents entries with status: "running", e.g. Claude Agent launches that returned "Async agent launched" and outlive the turn. Disposing the session kills the CLI and those agents. finishRunningSubagents then marks them done, so the user silently loses the work. This happens to any thread that delegates to background agents while 3+ other threads go idle.

Fix it in one place so all callers are covered:

• Never dispose a session while its thread has a running subagent. Skip it in the cap sweep, and don't arm (or re-arm on fire) the 30-min timer while one runs.
• Re-schedule once they settle. Subagent rows settle via ingestTaskNotifications and finishRunningSubagents, which already push threads:changed after PR Keep threads with running background subagents on the Working shelf #1423. When the last running subagent settles, call scheduleClaudeIdleReap again so the session can be reaped normally.
• Bound the exemption with a staleness cap (e.g. 2 h since the subagent row started, or since the session went idle), so a lost task-notification can't pin a session forever.

Write tests in electron/test/claude.test.js using the existing fake-claude scenarios (see "subagent-after-result", which keeps a background agent running after the turn's result):

1. With CLAUDE_IDLE_MAX exceeded, a session with a running background subagent is not disposed, while an idle one without subagents is.
2. After the subagent's task-notification settles it, the session becomes reapable again.

Run the tests with env -u CODER_GUARDRAILS_PATH node --test electron/test/claude.test.js. Before running the electron suite in a fresh worktree, clone node_modules and build core first. Then open a PR.

Filed from a Solenta suggested-work chip.